Hi Ioi,
On 2016-05-04 16:05, Ioi Lam wrote:
Hi Claes,
The HotSpot changes look good to me.
thanks!
I am not knowledgeable enough to comment on the top repo makefile
changes.
Regarding the JDK changes, in HelloClasslist.java:
Maybe add a comment about how/why you choose this particular set of
operations? When the JDK evolves in the future, how should this file
be changed?
I've done a few cycles of experiments to get something that shows a good
improvement on our minimal and server-style startup test, while
mitigating the regressions we'll inevitably see
on some of the GUI apps.
It'd definitely be interesting to iterate a bit on this as the JDK
evolves, and I think one of the benefits of this solution is that such
experiments are quite straightforward to setup.
I notice that no GUI classes are used. Is the reason (a) GUI classes
are not important anymore, or (b) the build would fail in a headless
environment if GUI classes were used by HelloClasslist.java?
A bit of both:
(b) Yes, requiring build environments to have a graphical environment is
complicated, but not impossible. We'd prefer not to have add that
requirement.
(a) Yes and no... while GUI classes aren't unimportant in and off
themselves, jigsaw, java.lang.invoke and other features are adding to
the number of java classes needed to bootstrap Java, and thus we'd see
the CDS archive grow ever larger if we simply kept refreshing the
classlists using the same technique as before. This actually has other
drawbacks such as reducing the chance the CDS archive will be
successfully mapped, especially on 32-bit platforms with ASLR. For
example on Windows 32-bit startup tests, the regression on GUI apps is
mostly offset by having a higher success rate.
Thanks!
/Claes
Thanks
- Ioi
On 5/4/16 6:36 AM, Claes Redestad wrote:
Hi,
please review this change to generate classlists at build-time
bug: https://bugs.openjdk.java.net/browse/JDK-8150044
webrevs:
top: http://cr.openjdk.java.net/~redestad/8150044/top.01/
jdk: http://cr.openjdk.java.net/~redestad/8150044/jdk.01/
hotspot: http://cr.openjdk.java.net/~redestad/8150044/hotspot.01/
The implementation generates an interim image consisting of a minimal
set of modules, then use this to run a small generator program to
load common utilities and facilities and dump the result of this to a
classlist that is then bundled with the final images.
The smaller number of classes on the default classlist (~1100 instead
of ~2500) requires some adjustment to the metaspace defaults.
This achieves the following:
- Removes a manual, error-prone process to update the versioned
classlists
- Ensures the classlists shipped with the JDK/JRE is up to date
with recent JDK changes, e.g., when moving classes from sun.* to
jdk.internal.*
- Automatically picks up and incorporates the output of jlink plugins
such as GenerateJLIClassesPlugin into the classlist
- Supports cross-compilation build targets, although it runs using a
build JDK that can run on the host platform to generate such
classlists (this isn't ideal, but no worse than the current
situation, where the versioned classlist for the host platform is
simply copied to the cross-compiled target)
There are a few concerns/drawbacks:
- It does add complexity to the build, and concern has been voiced that
this would adversely affect build times. However, I'm happy to say
that on my machine build times are roughly the same:
Before:
real 2m37.303s
user 35m33.576s
sys 3m46.476s
After:
real 2m36.168s
user 35m31.232s
sys 3m52.268s
(real time varies ± 5s from build to build)
- Startup on the specific applications we've used to generate the
classlists for previously suffer small regressions. These are
specifically rather dated AWT and Swing-based applications. OTOH,
startup characteristics generally improve on other applications
(minimal VM, jetty, etc...)
Testing: JPRT -testset hotspot
Thanks!
/Claes